The Age of Restricted AI
A striking shift is underway in the artificial intelligence industry: models are becoming so capable that their creators are choosing not to release them to the public. The latest example is Claude Mythos Preview, a model developed by Anthropic that can autonomously discover and exploit high-severity security vulnerabilities across major operating systems, web browsers, and critical infrastructure software. Rather than making it widely available, Anthropic has opted for a controlled rollout — a decision that may define how the most powerful AI systems are governed going forward.
Project Glass Wing: A Defensive Alliance
Instead of a public launch, Anthropic has funneled access to Claude Mythos through an initiative called Project Glass Wing. This program brings together more than 40 organizations — including Apple, Google, Microsoft, Amazon Web Services, CrowdStrike, Nvidia, and Cisco — entities collectively responsible for securing much of the world's digital infrastructure.
The rationale is straightforward: if a model can find and exploit critical flaws at machine speed, it is far better to place it in the hands of defenders first. The goal is to give these organizations a head start, allowing them to identify and patch weaknesses before malicious actors can take advantage of the same capabilities.
Thousands of Hidden Flaws Uncovered
The results so far are sobering. Claude Mythos has already uncovered thousands of previously unknown vulnerabilities, including decades-old bugs that traditional security tools and human auditors had missed entirely. This raises an uncomfortable question: if a single AI model can surface this volume of hidden flaws, how quickly could a hostile actor — armed with similar technology — weaponize them?
The sheer scale of discovery underscores both the promise and the peril of advanced AI in cybersecurity. On one hand, defenders gain an extraordinarily powerful tool for hardening systems. On the other, the existence of such a tool confirms that the attack surface of modern software is far larger than most organizations have assumed.
Access as a Safeguard
Perhaps the most significant implication of this story is not the technology itself but the precedent it sets. By keeping Claude Mythos off-limits to the general public, Anthropic is signaling that as AI grows more powerful, access itself becomes one of the most important safeguards.
This is a departure from the prevailing ethos of open release that has characterized much of the AI field. It suggests a future in which the most capable models are not products to be sold broadly but strategic assets to be deployed selectively — governed less by market demand and more by the potential for misuse.
Whether this model of restricted access can hold as AI capabilities continue to accelerate remains an open question. But for now, the message is clear: some tools are too sharp to hand out freely, and the organizations building them are beginning to act accordingly.